Relational Priors Increase Agreement but Not Accuracy in LLM-Based Multi-Agent Systems, Study Finds
A new arXiv preprint investigates the impact of explicit relational priors—such as trust or deference—on large language model (LLM) multi-agent systems. The study finds that these priors mainly act as convergence pressure, increasing coordination and agreement among agents, but do not reliably improve accuracy in objective tasks. The authors recommend against using relational priors by default, suggesting they be applied diagnostically and with careful monitoring of correctness when accuracy is important.
Why it matters: The findings caution that relational priors can boost agreement without improving correctness, informing safer design choices for multi-agent LLM systems.
Full story at: arXiv Computation and Language ↗